EP1484719A2

EP1484719A2 - Currency validation

Info

Publication number: EP1484719A2
Application number: EP04252403A
Authority: EP
Inventors: Gary Alexander Ross; Mark Andrew Girolami; Chao He
Original assignee: NCR International Inc
Current assignee: NCR International Inc
Priority date: 2003-06-06
Filing date: 2004-04-24
Publication date: 2004-12-08
Anticipated expiration: 2024-04-24
Also published as: GB0313002D0; US20040247169A1; EP1484719B1; EP1484719A3; US7639858B2

Abstract

A method for developing a template of a reference document, such as a bank note and using that template to validate other test documents. The method involves using images of a plurality of reference documents, such as banknotes and segmenting each image in a like manner into a plurality of segments (10, 12). Each segment is then classified using a one-class classifier to determine a reference classification parameter (14). These parameters are used to define a threshold reference classification parameter. Optionally, the step of segmenting is optimised to determine a preferred number of segments, and a preferred combination of these for use in a subsequent validation process. Preferably, this is done using a stochastic algorithm such as a genetic algorithm.

Description

The present invention relates to a method and system for devising a template for validating currency. The invention also relates to a method and system, in particular a self-service system such as an automated teller machine, for automatically validating or authenticating currency.

Banks and financial institutions have an increasing need to provide their customers with an expanding range of high standard financial self-services. The most popular and well-established service currently available is the Automatic Teller Machine (ATM). At present, ATMs are primarily used to allow customers to access personal financial information and to dispense cash. However, there is an increasing requirement to provide secure deposit transaction capabilities to complement these traditional ATM services.

Although some ATMs that can accept cheque and banknote deposits are available, only a limited number of these allow the validation and confirmation of the authenticity of the enclosed currency or cheques to be completed by the ATM. Instead, more generally, bank personnel have to manually confirm the validity of the deposit before completing the transaction and, in the case of a cash deposit, authorizing the re-circulation of the banknotes. For cash deposits, this is a problem, because it means that there is often a significant delay between receipt of the banknotes at the ATM and re-circulation of these notes.

Automated currency verification and validation systems for incorporation in ATMs have been proposed, and indeed some are currently available. The aim of these is to allow deposited banknotes to be verified and validated within an ATM and then re-circulated as part of the withdrawal service of that same ATM. However, a fundamental danger of automated currency validation is the possibility of utilizing a currency recycling facility within ATMs as a means of circulating counterfeit currency.

At present, in order to validate currency classical pattern recognition techniques are used. For these to be effective, a representative sample of both genuine and counterfeit currency has to be available to devise discriminator templates. All other notes are compared with these discriminator templates. Notes that are identified as genuine are accepted. All other notes, such as counterfeits or defaced genuine notes or unknown notes, are rejected. In current systems, the templates are prepared manually by the observation of collected data and by putting more emphasis on some of the data than others. However, this handcrafted template approach is very time consuming and costly, and prevents rapid reaction to new counterfeits. Manufacture of the templates is done centrally, which requires the movement of data or notes and the use of an expert craftsman to carry out the work. This means that the templates can take several weeks to prepare, which in turn means that there is a delay between issuance of new currency, and the inevitable new counterfeits, and a validation mechanism for banks.

An object of the invention is to provide method for devising a template for a document such as a cheque or banknote that has to be validated.

Another object of the invention is to provide an improved method for verifying and validating such documents, and in particular banknotes.

According to a first aspect of the present invention, there is a provided a method for creating a template for a document, such as a banknote, that has to be validated, the method comprising using a scanner or imager to capture images of a plurality of reference documents, such as banknotes, and using the captured images to define a reference template. By template it is meant a representation of the reference images, which typically is a numerical or statistical representation of those images.

By using a plurality of scanned images of genuine documents to define a template, there is provided a very simple and effective method for dynamically creating a reference template that can be used in later validation processes. Because the template is built up merely by scanning images of genuine notes, there is no need for expert intervention in the manufacture of the template. This means that reference templates can be generated quickly and cheaply when new currency is introduced.

The imager or scanner may be an optical or thermal or magnetic imager or scanner. As one example, the imager may comprise a plurality of light sensors, preferably each being sensitive to a different colour. In this case, the image may be represented by colour differences. Here, the template would be a representation of the distribution of optically measured colour differences within the inks that cover the document. This is particularly useful for banknotes. In a preferred embodiment, four different sensors are provided and the image is represented by a six dimensional vector, each dimension of the vector being indicative of differences in intensity between signals received at two of the four sensors. The imager or scanner could be included in a self-service terminal so that the template can be determined locally as and when desired. Alternatively, the imager or scanner could be provided in a secure area of a bank for use by bank personnel.

The method may further involve segmenting each image in a like manner into a plurality of segments, and using like segments of the images to determine a reference segment or reference segment parameter for each segment. By using or comparing like segments of each of a plurality of images of genuine documents, it is possible to build up a segmented reference image that can be used in a subsequent validation technique.

Preferably, the step of using the segments of the images involves determining a reference classification parameter for each segment; and defining a threshold for the reference classification parameter. By defining a threshold value for the reference classification parameter or range of such parameters for each segment of the genuine note, it is possible to determine whether other test notes of unknown status are genuine by determining classification parameters for segments of the test note and comparing these with the threshold for the reference classification parameters. An advantage of doing this is that by varying the threshold value for the classification parameter, there is provided a very simple way for changing the acceptable margin of error. In the banking environment, this is useful, because it allows banks flexibility in setting how many counterfeit notes they would be prepared to accept and likewise how many valid notes they would be prepared to reject.

Preferably, the step of classifying comprises using a one-class classifier. The one-class classifier may use a parametric D² test, as described in the book "Multivariate Statistical Methods" (third edition), by Morrison, McGraw-Hill Publishing Company, New York (1990). Another one-class classifier that could be used is a semi-parametric test based on a mixture of Guassians. This semi-parametric test may employ a bootstrap.

Preferably, the method further involves optimising the step of segmenting, for example optimising the number of segments. This may be done using a stochastic algorithm such as a genetic algorithm. The step of optimising may involve determining a preferred combination of segments that could be used in a validation process, which preferred combination is best for providing an indication of validity.

The method in which the first aspect of the invention is embodied is preferably implemented using software. To this end, according to another aspect of the invention, there is provided a computer program, preferably on a data carrier or computer readable medium, having code or instructions for using images of a plurality of reference documents, such as banknotes, which images are captured using a scanner or imager, and processing the plurality of images to determine a reference template.

Preferably, the computer program has code or instructions for segmenting each image in a like manner into a plurality of segments, and comparing like segments of the plurality of images to determine a reference segment image or reference segment parameter for each segment.

The code or instructions for comparing the segments of the images may be operable to classify each segment to determine a reference classification parameter for each segment, and preferably a threshold for the reference classification parameter. The code or instruction for classifying may implement a one-class classifier, for example a one-class classifier that uses a parametric D² test or a semi-parametric test based on a mixture of Guassians, preferably employing a bootstrap.

The computer program may comprise code or instructions for optimising the step of segmenting, preferably by determining a preferred number of segments and/or a preferred combination of segments for use in a validation process. The code or instructions for optimising the step of segmenting may use a stochastic algorithm.

According to yet another aspect of the present invention, there is provided a system for developing a template for a document that has to be validated, the system comprising means for using images of a plurality of genuine documents such as banknotes; means for segmenting each image in a like manner into a plurality of segments; means for classifying the segments of the images to determine a reference classification parameter for each segment; and means for defining a threshold for the reference classification parameter.

Once the template is dynamically defined using any of the techniques in which the preceding aspects of the invention are embodied, it can be used in a subsequent validation process. To this end, according to still another aspect of the invention, there is provided a method for verifying or validating documents comprising capturing an image of the document; segmenting that captured image into a plurality of different segments; comparing one or more of the segments of the captured image with corresponding segments of a reference image, and determining whether the document is valid or invalid based on the results of the step of comparing.

By comparing one or more of the segments of the captured image with the corresponding segments of the reference image, it is meant comparing the image segments or any data associated with or indicative of those image segments. In particular, the step of comparing the segments of the image with the corresponding segments of the reference image may involve classifying each segment to determine a classification parameter, and comparing that classification parameter with a threshold reference classification parameter for the corresponding segment of the reference image. The step of classifying may use a one-class classifier. The one-class classifier may use a parametric D² test, or a semi-parametric test based on a mixture of Guassians. The semi-parametric test may employ a bootstrap.

The method may involve comparing a plurality of the captured image segments with a corresponding plurality of the reference segments, thereby to determine the validity of each of these segments. To determine the overall validity of the document, the validity of a plurality of the segments may be combined. This may be done using a unanimous vote, in which all of the segments considered have to be valid in order to give an overall result that the document is valid. Alternatively, a majority vote be used, in which only a majority of the segments considered have to be valid in order to give an overall result that the document is valid.

In order to combine the results of the validation of each segment, a figure of merit may be assigned to each segment, the figure of merit being indicative of whether or not a segment is valid. The figure of merit may be an integer. As a specific example the integer 1 may be used to indicate that a segment is valid and value 0 may be allocated to segments that are invalid. To determine overall validity, the figures of merit for each of the plurality of segments may be multiplied together.

When a document is determined to be valid, the captured image or any data associated with it may be used to up-date the reference image. In this way, the reference image is continuously being improved.

According to yet another aspect of the invention, there is provided a system, preferably a self-service terminal, such as an ATM, for verifying or validating documents, the system comprising means for capturing an image of the document; means for segmenting that captured image into a plurality of different segments; means for comparing one or more of the segments of the captured image with corresponding segments of a reference image, and means for determining whether the document is valid or invalid based on the results of the step of comparing.

The system may comprise means for using the captured image or any data associated with it to up-date the reference image, if it is determined that the captured image is valid.

According to still another aspect of the invention, there is provided a method for creating a template of a document, such as a banknote, comprising capturing an image of the document, classifying the image using a one-class classifier, and using the classification data as a reference or template for use in a subsequent validation process.

According to yet another aspect of the invention, there is provided a method for verifying or validating documents comprising capturing an image of the document, classifying the image using a one-class classifier, comparing the classification data with reference classification data and determining whether the document is valid or invalid based on the results of the step of comparing.

Various aspects of the invention will now be described by way of example only and with reference to the accompanying drawings, of which:

Figure 1 is a flow diagram of the steps taken to create a template of a banknote, or other image rich document;

Figure 2 is a block diagram of a system for creating a template of a banknote;

Figure 3 is a representation of a portion of a banknote;

Figure 4 is a representation of the banknote of Figure 1 that is segmented into four segments;

Figure 5 is a table showing false negative and false positive rates for two different one-class classifiers;

Figure 6 is a table showing false negative and false positive rates for a two-class classifier;

Figure 7 is a table showing results of an optimised note segmentation and classifier combination;

Figure 8 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 7 for the search space (3,3);

Figure 9 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 7 for the search space (7,7);

Figure 10 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 7 for the search space (15,15);

Figure 11 is a table of results on an optimised note segmentation and classifier combination for another one-class classifier;

Figure 12 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 11 for the search space (3,3);

Figure 13 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 11 for the search space (7,7);

Figure 14 is a diagrammatic representation of an optimised note segmentation and classifier combination for the results of Figure 11 for the search space (15,15);

Figure 15 shows data indicating the variability of a genetic algorithm that is used in a process for optimising the segmentation and classifier combination, and

Figure 16 is a block diagram of an ATM that is adapted to create a template of a banknote and/or automatically validate currency deposited therein.

The present invention provides a system and method for dynamically determining a reference template for a banknote using genuine notes and using that reference template to verify or authenticate banknotes deposited by customers. By providing a mechanism for dynamically generating the reference template, when new currency is issued, a reference template can be readily obtained merely by using examples of genuine notes. This means that banks can be more responsive to changes in currency. This closes a window of opportunity for fraudsters. In addition, if a representative number of counterfeit examples are available, an additional template for these can also be dynamically generated, so that notes deposited by customers can be compared with one or both of the reference template for genuine notes and the counterfeit template.

Figure 1 shows the steps taken to create a template of a banknote in accordance with the invention. As a first step, to characterize banknotes to build up a suitable template for use in a later validation scheme, images of genuine notes are captured 10. To do this, a plurality of genuine or reference notes 24 is scanned or imaged using a scanner/detector arrangement 20 and the resultant image data is processed using a suitable processor 22, see Figure 2. The step of imaging can be done in any suitable manner, but preferably by using a set of narrow-band sensors, which provide responses to radiation reflected from across the surface. In a preferred example, four sensors are used. The sensors used are all light emitting diodes (LEDs). Each emits a different type of radiation. Preferably, the sensors comprise a red LED, a green LED, a blue LED and an IR LED. Light from these is directed onto the surface of the banknote and four suitable detectors are provided to detect the reflected signals. In this way, four separate images are captured, one for each of the four LEDs.

Each of the images captured by the various detectors has a plurality of pixels, which represents the intensity of the radiation received by the detectors at a given point. The image data for each banknote or a selected region of each banknote, is segmented into nonoverlapping and equal-sized sub-regions, see step 12 of Figure 1. This is done by over-laying a pre-determined grid on the captured image. Pixels that fall within the same portion of the grid are taken to define a particular segment of the image. As a specific example, Figure 3 shows a selected region of a Bank of Scotland £10 note. Figure 4 shows a segmentation of the selected region of the bank note of Figure 3 into four equal sized square segments. For the purposes of illustration, a one-dimensional distribution for each region is also shown. More generally, each note is segmented into R by C sub-regions, which can be numbered from 1 to L, left-to-right and then top-to-bottom wise, where R and C are respectively the number of segmented rows and columns, and L = RC is the total number of segmented sub-regions.

Every pixel of each of the four images is compared with the corresponding pixel of each of the other images, and the difference in intensity is stored. In particular, the difference in intensities between the red and green data is determined, the difference in intensities between the red and blue data is determined etc. In all a six-dimensional feature set is derived from the responses from the four detectors for every segment of the image, so that every segment is represented by such a six-dimensional feature set.

To classify the document, individual classifiers on the feature sets extracted from each segment or region are devised. Each segment is individually considered in order that it can be classified, see step 14 of Figure 1. In the present case, a mathematical function termed a one-class classifier is used. This type of classification is useful where information on only one class is available. It defines a boundary around the known class, so that samples that fall outside this boundary are deemed not to belong to the class. In the present example, the information that is available is the pattern and layout of the segments of genuine banknotes. Each of these is classified to define suitable boundaries. The resultant classifications for each segment are then used in order to create a reference template, see step 18 of Figure 1, which can be used to validate or authenticate other banknotes. As part of the validation process, different segments may be given different levels of significance or weights. This is because certain segments of a banknote may be more difficult for fraudsters to copy and so may be more likely to indicate the presence of a counterfeit. This will be described in more detail later.

One-class classifiers are well known. Examples of these are described in "Support vector domain description" by Tax et al in Pattern Recognition Letters, 20(11-13) 1999 1191-1199; "One-class classification" by Tax, Technische Universiteit Delft Ph.D. Thesis, (2001) and "Support Vector Novelty Detection Applied to Jet Engine Vibration Spectra" by Hayton et al in Advances in Neural Information processing Systems 13, MIT Press, 946 - 952, (2001). One example of a preferred classifier is based on a parametric D² test, as described in the book "Multivariate Statistical Methods" (third edition), by Morrison, McGraw-Hill Publishing Company, New York (1990). Another example of a suitable one-class classifier uses a semi-parametric test based on a mixture of Guassians, which semi-parametric test employs a bootstrap as described in "An Introduction to the Bootstrap" by Efron et al Chapman Hall/CRC Press LLC, Boca Raton, Florida (1998). Each of these will be described in more detail.

A specific example of a classifier that can be used to devise a template and subsequently be used to validate currency will now be described. Consider N independent and identically distributed p-dimensional vector samples, in this case, the feature set for each segment of each banknote can be represented by x₁, ..., x _N ∈ C with an underlying density function with parameters  given as p(x|). The following hypothesis test is given for a new point x_N+1 such that H ₀ : x_N+1 ∈ C vs H ₁ : x _N ₊₁ ∉ C, where C denotes the region where the null hypothesis is true and is defined by p(x|) and the significance level of the test. Assuming that the distribution under the alternate hypothesis is uniform, then the standard log-likelihood ratio for the null and alternate hypothesis can be employed as a test statistic for the null-hypothesis. This test can be expressed as follows:

This can be used as a statistic for the validation of a newly presented note. More specifically, a reference set of notes N is scanned and the six dimensional feature vectors are analysed to determine a threshold value for λ for each segment of the note. The threshold value defines a boundary between what is acceptable and what is unacceptable. In practice, this threshold value defines a lower limit. If the value of the test statistic λ for a segment of test sample were below this, the test segment would be rejected. Likewise, if the value of the test statistic λ for a segment of test sample were above this, the segment would be accepted. Hence, in order to validate a test segment, it is scanned, segmented in a like manner to the segmentation used to define the reference template and a value for λ is determined for each segment. These values are compared with the reference threshold values for the corresponding segments to determine the validity of each of the segments.

One specific approach is to assume that the feature vectors for the banknote have a multi-variate Gaussian distribution. A test can then be applied to assess whether each point in a sample shares a common mean. Examples of this type of test are described in more detail in the book "Multivariate Statistical Methods" (third edition), by Morrison, McGraw-Hill Publishing Company, New York (1990). As a specific example of a suitable test, consider N independent and identically distributed p-dimensional vector samples x₁, ..., x _N from a multi-variate normal distribution with mean µ and covariance C, whose sample estimates are µ and_N and C and_N. From this sample consider a random selection denoted as x₀. For this selection the associated squares Mahalanobis distance is: D2 = (x 0 - µ N ) T C -1 N (x 0 - µ N ) This can be shown to be distributed as a central F-distribution with p and N-p-1 degrees of freedom by: F = (N - p - 1)ND 2 p(N - 1)2 - NpD 2 Then, the null-hypothesis of a common population mean vector for x₀ and the remaining x _i will be rejected if: F> Fα;p,N-p-1 where α is a pre-set significance level for the test and F_α,p,N-p-1 is the upper α100% point of the F-distribution with (p, N - p - 1) degrees of freedom, so that by using this threshold α times 100% genuine samples will be rejected. The incremental estimates of the mean and covariance µ and_N and C and_N respectively are as follows: µ N+1 = 1 N + 1 {Nµ N + xN + 1} C N+1 N N + 1 C N + N (N + 1)2 (xN +1 - µ N )(x N+1 - µ N ) T By using the above expressions for the estimates of the mean and covariance and the matrix inversion lemma, the equation D ² for an N-sample reference set and an (N + 1)th test point becomes: D 2 = δ T N+1 C -1 N+1δ N+1 where δ N+1 = N N + 1 (x N+1 - µ N) and C -1 N+1 = N + 1 N C -1 N - C -1 N δ N+1δ T N+1 C -1 N N + N N + 1 δ T N+1 C -1 N δ N+1 Hence,

where D² _N+1,N is δ^T _N+1 C and_N ^-1 _δN+1. Hence, a new point x_N+1 can be tested against an estimated and assumed normal distribution for a common estimated mean and covariance. It should be noted that employing the log-likelihood ratio for the one-class hypothesis test, the test statistic can be derived directly.

In practice, when the parametric D² test is used, N reference notes are scanned and segmented. Each of these segments is classified by determining a value for D² for every one of the reference notes and using this to set a reference threshold value for each segment. This threshold value is used to determine whether corresponding segments of subsequent test notes are valid.

The above analysis is based on the assumption that the feature vectors are distributed as multivariate Gaussian. Often this does not hold in practice, although may be an appropriate pragmatic choice in many applications. However, this assumption can be relaxed and arbitrary densities can be considered. The density under a mixture model has the following standard form:

where P (j), j = 1, ..., M are the mixing parameters. These are chosen to satisfy the constraints

and P(j) ≥ 0. The component density functions p(x|j) are normalized so that ∫p(x|j)dx=1. As a specific example, Gaussian mixture models are used, so that:

The required parameters can be estimated using the Expectation Maximization (EM) algorithm. A technique for doing this is described in "Neural Networks for Pattern Recognition" by Bishop, Oxford University Press, New York (1995). This density can then be employed in computing the log-likelihood ratio. Unlike the case for the multivariate Gaussian distribution there is no analytic distribution for the test statistic λ under the null-hypothesis. To obtain the otherwise non-analytic null distribution under the mixture of Gaussian density, bootstrap methods can be employed. By doing this, the various critical values of λ_crit can be established from the empirical distribution obtained. It can be shown that in the limit as N tends to infinity, the likelihood ratio can be estimated by the following: λ = sup ∈⊖ L 0()sup ∈⊖ L 1() → p(xN +1; N ) where p(x_N+1;  and_N ) denotes the probability density of x_N+1 under the model estimated by the original N samples. After generating B bootstrap samples from the reference data set and using each of these to estimate the parameters of the mixture distribution  andⁱ _N, B bootstrap replicates of the test statistic λⁱ _crit (i = 1,..B) can be obtained by randomly selecting an (N + 1) th sample and computing p(x _N+1;  and i / N) ≈ λ i / crit. By ordering λⁱ _crit in ascending order, the critical value α can be defined to reject the null-hypothesis at the desired threshold or significance level if λ is less than or equal to λ_α, where λ_α is the jth smallest value of λⁱ _crit and α = j/(B+1). By scanning N different notes; segmenting the images and calculating λ for each segment to determine a reference threshold, it is possible to validate subsequent notes by determining whether one or more of the test segments is within the tolerance level of what is acceptable.

Any of the above classifiers could be used to validate banknotes or other image rich documents. In any case, when the required one-class classifier is selected, it is trained on each segment or sub-region of the note, so that boundary conditions for each segment can be defined, that is limits that define what is regarded as being an acceptable or unacceptable variation from the reference segment. The classifier for the ith sub-region is denoted as D_i, where i = 1,...,L. The classifier is trained using feature matrices for the reference image, where these reference feature matrices can be represented by X_k = (x_i ^k), where k = 1,..N; i = 1,..L, and x_i ^k = [x₁ ^k, x₂ ^k, ..x_p ^k]^T is a p-dimensional feature vector of the kth sample in the ith sub-region. Once trained on the reference notes, the classifiers for the various segments can be used to validate other banknotes. In particular, the classifiers D_i can be used to determine whether or not each segment of the note is within the defined classification boundary and so acceptable or outside the defined boundary and so rejected.

When a test sample t is presented, its feature matrix T = t_i, where i = 1,.., L, is extracted. Then, using the ith one-class classifier D_i trained by {x_i ¹, x_i ²,..x_i ^N}, t_i can be tested using the same rule to give the decision y_i (t) whether it is rejected or accepted. As a specific example y_i(t) may be defined as 0 if t_i is rejected by Di and 1 if t_i is accepted by D_i. There are various ways in which the final decision Y(t) for the test sample t can be calculated. As one example, the product Y(t) of individual decisions y_i(t) can be determined. The test note t is rejected as a forgery if Y (t) = 0, or accepted as a genuine note if Y (t) = 1. This product combination decision rule described above can be seen to be equivalent to a unanimous vote. This means that a note is only accepted if all selected classifiers indicate that it is genuine, otherwise it would be rejected. Unanimous voting is an extreme situation of commonly used majority voting when requiring all the voters to agree with consensus. Unanimous voting is preferred technique here, because of the higher cost assigned to incorrectly accepting forgeries. Nevertheless, in some circumstances majority voting may be useful.

To illustrate the differences between unanimous and majority voting techniques, consider the following definitions for the final decision Y_MV(t) made by employing a majority vote and the final decision Y_UV(t) by an unanimous vote for a note under test t:

Here, INT (·) means rounding down the element to its nearest integer, and both Y_MV (t) and Y_UV (t) can only have the value of either 1 or 0. The expression for Y_UV (t) can be re-written as:

Given N_f counterfeit samples {t _f ^k1}, where k₁ = 1, ..N_f, and N_G genuine samples to test {t_g ^k2}, where k₂ = 1,.., N_g, the False Negative (FN: false acceptance rate of forgeries) and False Positive (FP: false rejection rate of genuine notes) can be calculated by

where Y( ) can be either Y_uv(t) or Y_MV(t) . As can be shown from the above equations, the unanimous vote approach achieves lower FN than the majority vote. However, at the same time the unanimous vote also produces higher FP than majority voting. This is because majority voting looks at the average information rather than the individual characteristics. Therefore, for testing genuine notes, as their features are relatively uniform within the reference samples in all sub-regions, majority voting can potentially achieve better results. In contrast, for identifying counterfeits, a unanimous vote may be more appropriate. This is because very high quality counterfeits have similar feature distributions as genuine notes in a number of sub-regions and differences exist in a few specific sub-regions where the genuine features might be too complex to be completely duplicated by the counterfeiting process. By taking account of the average information, majority voting may give the final decision according to most of the sub-regions that might give an incorrect judgement for counterfeits. Hence, for identifying counterfeits unanimous voting may be preferred.

Whilst the unanimous vote based technique does not have the problem of giving wrong decisions towards counterfeits, it does however suffer from falsely rejecting, for example, poor quality genuine notes due to the worn nature of the note in some sub-regions. Hence, as with all statistical tests, there is a trade-off between FN and FP. In this particular application, FN is more important than FP. The proposed log-likelihood ratio tests described previously allow the expected FP level of individual classifiers to be specified by setting the significance levels of each test. Therefore, the extreme unanimous vote is preferred to balance the overall FN and FP performance. As an example, by segmenting a whole note into 3 by 3 equal sized regions and combining all 9 classifiers, in testing 1000 genuine notes and 1000 forgeries, majority voting achieved FN=23.9% and FP=0.50%; while unanimous voting achieved values of FN=2.30% and FP=8.20%.

When applying the unanimous vote combination rule, selecting appropriate classifiers to be combined is important. Not all the classifiers built on sub-regions have to be combined and indeed doing so may reduce the robustness of the whole classification system as mentioned above. This is because some of the segments of the note may be more difficult to copy than others and so may be more likely to provide evidence of a counterfeit. Hence, to make the classification process more accurate, there is provided an optimisation technique for identifying the ideal number of segments and/or the combination of segments that is most likely to indicate that a segment is not valid and so potential counterfeit activity.

Considering an R _max×C _max searching space where R _max and C _max are preset maximum numbers of rows and columns of the segmentation, the total number of possible ways to segment and combine the individual regions would be:

where

is the number of combinations of ij regions taken k at a time. If R _max = 15 and C_max = 15, P = 5.3923×10⁶⁷. This value exponentially increases following the changes of bigger R _max and C _max. Therefore, it is impossible to carry out exhaustive searching experiments. To achieve a useful result an optimised note segmentation and classifier combination strategy is proposed to gain the best trade-off between FN and FP performance can be achieved. In this, an optimal number of segments and an optimised classifier combination set are identified by a stochastic genetic algorithm (GA). This is useful because, as will be appreciated, different currencies have different feature distributions, so it is impossible to define a uniform segmentation and combination that is suitable for all possible currencies.

To optimise the note segmentations as well as the classifier combination at the same time, a stochastic optimisation algorithm is used, in particular a genetic algorithm. This has one chromosome c, which is composed of three gene groups c_r, c_c and c_d, where c_r = [r ₁, r ₂, ...,r _n1], C_c = [c ₁ , c ₂, ..., c _n2] and c_d = [d ₁, d ₂, ..., d _n3] are all binary strings. The first two gene groups c_r and c_c respectively represent the number of rows and columns for a note segmentation. Their relationships with R and C are given by: R = BinToDec (C_r) and C = BinToDec (c_c) , where "BinToDec" means transforming a number from a binary expression into decimal expression. Given the number of maximum rows R _max and maximum columns C _max which define the size of the intended segmentation search space, c_r's length n₁ equals the length of the binary string expression of R _max; and c_c's length n ₂ equals the length of the binary string expression of C _max. The third gene group c_d represents the combination of classifiers. Its elements d_i and length n ₃ are defined as d_i =0 if D_i ∉ S and d_i =1 if D_i ∈ S where i=1,2,...,n₃; n₃=R_maxC_max and S is the optimised classifier combination set. For example, given R _max = 3 and C _max = 3, then n ₁, n ₂ and n ₃ would respectively equal to 2, 2 and 9, and a chromosome

means segmenting a note into 3×2 (i.e. R = 3, C = 2) and combining the classifiers D₁, D₂, D₄ and D₆ respectively built on the 1^st, 2^nd, 4^th and 6^th sub-regions.

The GA for optimising the system can be summarised as follows: (1) Initialize a random population of n chromosomes; (2) Perform crossover and mutation operations to create another n offspring; (3) carry out the validation operation, and calculate a fitness for each chromosome; (4) Select n chromosomes with the best fitness among all the parents and offspring as the next generation, and (5) if convergent or over the preset maximum iteration steps then stop, otherwise go to step (2). To implement this overall strategy, the GA has three gene operations for the chromosome. The first operation is gene crossover. Considering the characteristics of the chromosomes, two crossover points in one chromosome are set to complete the crossover operation with other chromosomes. One is the point between c_r and c_c; and the other is the point in the middle of c_d. For example, if there are two chromosomes c₁ and c₂ represented as follows:

after crossover, two new chromosomes C₁ ^new and C₂ ^new are produced, where

The second GA operation is gene mutation. For this, a standard gene mutation operation is used, i.e. every time randomly selecting a gene (bit) to mutate from 1 to 0 or from 0 to 1. The third GA operation is validation. Because once the c_r and c_c have changed, only the first RC genes in c_d are effective, it is necessary to clean up the ineffective genes in order to avoid them affecting other chromosomes in further operations. Therefore the validation operation is designed to set all the n ₃ -RC ineffective genes in c _d to be zeros.

Once the validation operation is carried out, a fitness function is calculated by balancing the performance of FN and FP in the validation set with the GA. Given N ₁ genuine samples G = {g_k1}. where k₁ = 1, ...N₁ and N ₂ counterfeit samples F = {f_k2}, where k₂ = 1, ..N₂
where both g_k1 and f_k2 are p-dimensional column vectors, then the decisions y_i(g_k1) and y_i(f_k2) of the ith classifier Di towards validation samples g_k1 and f_k2 can be calculated. These are used to determine a fitness function, which is defined as:

where the first term indicates the True Negative (TN: = 1 - FP) performance and the second term indicates the True Positive (TP: = 1 - FN) performance. A chromosome that has a larger value of fitness function will have better performance when applying the segmentation and combination that this chromosome represents.

An alternative approach to evaluate the fitness of chromosomes is to use Receiver Operating Characteristics (ROC). This is described in the paper "Genetic programming for combining classifiers" by Langdon et al, Proceedings of the Genetic and Evolutionary Computation Conference (GECCO'2001), San Francisco, USA, Morgan Kaufmann (2001). However, this requires running tests for each individual sub-region by setting different significance levels. This would lead to higher computation costs.

For the procedure of selecting the segmentation and combination of regions described above, some validation samples in the opposite class (counterfeits) are needed to evaluate the performance of the selected segmentation and combination. However, compared with the amount of counterfeits that would be required to build a two-class classifier, small amounts of counterfeit validation samples suffice for this validation purpose.

In order to compare the effectiveness of the one-class classifiers proposed, various tests have been done on both whole notes and segmented notes. Using 3324 genuine notes as a reference set, based on a single feature vector defining the whole note, the D ² and Mixture of Gaussians (MoG)+Bootstrap tests were employed on 798 counterfeit notes by setting different critical values, that is test significance levels, α= 0.01, α = 0.05 and α = 0.10 to specify the expected FP rate for each test. To obtain the FP rate for genuine notes 10-fold cross-validation was employed. Test results for both tests are listed in Figure 3. It can be seen that the semi-parametric test for all significance levels provides significantly lower levels of false negatives (counterfeits accepted as genuine) than that obtained by the D ² tests.

For comparison simple binary classifiers or two-class classifiers have also been tested, specifically estimating the class conditional densities based on both a Gaussian and Mixture of Gaussian models. In this experiment, 3324 genuine and 798 counterfeit were used to train the two-class discriminators. Test results are shown in Figure 4, where 10-fold cross-validation is used for both classes to calculate the values of FN and FP. Employing simple binary classifiers can provide an improvement of around 4% in FN rates over the D ² test, whilst the difference in performance for the MoG+Bootstrap tests is less pronounced. However, employing the one-class classifiers it is possible to reduce the costly FN rate by increasing the significance level of the test and so increase the number of genuine notes rejected (FP). Clearly in cases where large numbers of examples of counterfeits are available then employing binary classifiers will be advantageous, however one-class classifiers are most useful where the numbers of examples from both classes is highly imbalanced such that the distribution of only one class can be reliably estimated.

The results set out in Figures 5 and 6 are based on an analysis of the entire note. To test the effect of segmenting the note various segmentations have been tried. In one example R_max and C _max were set to be (3,3), (7,7) and (15,15). Of course, it will be appreciated that R_max and C _max do not have to be identical they can be any positive integer greater than zero. Using these values optimised note segmentation and classifier combinations were determined, using individual D ² tests at 1% significance levels. In the experiments the GA used 50 chromosomes in each population. A sample of 2500 genuine notes was used as the reference set for the D ² test. The GA was trained using 10-fold cross-validation on the 2500 genuine samples as well as 300 counterfeit examples. Finally, tests were performed on an independent test set of 824 genuine notes and 498 counterfeits. The reported test results, shown in Figure 7, are the best selected from 200 repeated runs of the GA.

The optimised note segmentation and classifier combination strategies obtained by running the GA using the D ² test are shown in Figures 8 to 10. In comparison to the best results obtained by considering the note as a whole the segmentation and combination strategy has reduced the number or counterfeits wrongly accepted (FN) from 5.76% to 2.41% whilst also significantly reducing the number of genuine notes being erroneously rejected (FP) from 10.1% to 3.03%. Regions of both high contrast with little detail and regions with high levels of detail are combined. Intuitively it would seem likely that the high contrast regions help to reduce the number of genuine notes rejected whilst the highly detailed regions assist in reducing the number of counterfeits wrongly accepted - as these are the regions which are typically more difficult to copy successfully.

The test described above with reference to Figures 8 to 10 was repeated using the mixture of Guassian and bootstrap test. Test results (the best selected from 200 repeated runs) are shown in Figure 11. The optimized note segmentation and classifier combination strategies obtained by running the GA for the mixture of Guassian and bootstrap test are shown in Figures 12 to 14. In this case, when segmenting the note into 4×11 regions, the best performance for the training samples is achieved, whilst the best performance for the independent test set is achieved by segmenting the note into 7×3 regions. This indicates that a more complex segmentation will not necessarily achieve better performance on an independent test set and so a level of 'overfitting' can be observed. In comparison to the best D ² test results obtained, the best MoG+Bootstrap provides a lower FP value, i.e. number of genuine notes rejected, with a marginal increase in the number of FN's. The decrease in FP value is due to the reduced number of regions being tested (two) and so only two tests at the 1% significance level are being performed. In contrast three regions are combined in the D ² test and so the expected FP rate is approximately 3x the individual significance levels. Considering the segmentations and combinations again similar regions of low detail and high detail content have been selected as in the previous case.

Because the GA is a stochastic optimisation algorithm, it yields different results in different randomly initialised runs. The GA can be run N _r times and the note segmentation and classifier combination strategy that achieves the best performance can be selected. To investigate the level of solution variability, the GA using mixture of Gaussian and bootstrap was repeatedly run N _r = 200 times. For the search space (3,3), only two different results (segmentation and combination strategies) were obtained during 200 randomly initialised runs. These are shown in Figure 15. For the search space (7,7) thirty-four different segmentation and combination strategies were obtained from the 200 runs. Among them, the best strategy was a 7×3 segmentation of the note and then combining classifiers D₁₂ and D₁₃. This achieved a training performance: FN = 3.67%, FP = 1.88% and a testing performance: FN = 2.61%, FP = 1.46%. The inter-quartile range (IQR) of values of the training FN, FP and testing FN, FP for those 34 strategies over all runs are respectively: 1.67%, 1.04% and 1.61%, 3.03%. IQR computes the difference between the 75^th and the 25^tn percentiles of the sample. It is a robust estimate of the spread of the data, since changes in the upper and lower 25% of the data do not affect it. On observation the distribution of FN and FP is not normal, IQR is more representative than the standard deviation as an estimate of the spread of the body of the data. For the search space (15,15), one hundred and twenty-eight different strategies were obtained through the 200 runs. Among them, the best strategy is a 4×11 segmentation followed by the combination of classifiers D₁₈, D₂₂ and D24, which achieves training performance: FN - 0.67%, FP = 2.64% and testing performance: FN = 2.81%, FP = 2.79%. The IQR of training FN, FP and testing FN, FP of those 128 strategies over all runs are respectively: 3.50%, 1.10% and 1.61%, 1.70%.

In practice, the methods described above for defining a template for validating a bank note and using that template to validate test notes may be implemented within a self-service environment, such as an ATM. In this way, the ATM can be trained in a simple and effective manner to create a reference template and use that template to validate deposits in a secure manner. This can be done without the need for human intervention, so that the entire process is automatic.

Figure 16 shows an example of an automated teller machine that is configured to implement these techniques. This includes a housing 20 that has a front fascia 22 that has a screen 24 and keyboard 26. Defined in the fascia are three slots, a first one 28 of these being for receiving a user's bank card, the next being provided for dispensing cash through 30 and the third one 32 being for depositing cash into a secure store 34 in the housing 20. One or more carrier mechanisms (not shown) is provided for moving cash from the secure store to the dispensing slot, so that it can be dispensed therefrom, and for moving cash received through the deposit slot to the store. On a path 36 between the deposit slot 32 and the cash store 34 is an imager or scanner 38 for imaging deposited banknotes. In a preferred example, this is the four-sensor imager 38 previously described. The carrier mechanism is operable to carry a banknote received from the deposit slot 32 to the imager 38. Once appropriately positioned relative to the imager 38 one or more suitable images can be captured.

Connected to the imager 38 is a central processing unit 40 that includes software for analysing the captured images. This software has code or instructions for defining a template based on a plurality of reference images captured from deposited reference notes, in accordance with the techniques described previously. Additionally, the software has code or instructions for validating currency deposited into the ATM, after the initial template is built up, and using an image captured by the imager 38. Again, this validation is done in accordance with the techniques previously described. In the event that currency deposited is deemed to be valid, it is carried into the secure store 34, from where it can be dispensed. In addition, the customer's personal account is up-dated to show the deposited amount. In the event that currency deposited is deemed by the validation software to be invalid, it is returned to the customer either through the deposit slot 32 or the dispensing slot 30.

Whilst the ATM of Figure 16 is described as having software for defining a template as well as for validating currency deposited, this is not essential. Instead, the template may be devised at a remote scanning/imaging/processing system, using the method previously described, and the classification threshold parameters may be provided to the terminal. For example, the classification threshold parameters may be downloaded to individual terminals. Alternatively, where a plurality of ATMs are provided in a distributed network, the template information may be stored at a remote location and shared between the remote terminals.

The present invention provides a solution to the long-standing problem of how to validate currency in a self-service terminal. The preferred method is based on the segmentation of the whole note into different sub-regions followed by the combination of individual classifiers built on those regions. It has the advantage of being applicable to notes of any currency when there is insufficient numbers of counterfeit examples to build a robust binary classifier. In a preferred embodiment, one-class classifiers are employed and a genetic algorithm is used in selecting the optimised note segmentation and classifier combinations. The method in which the invention is embodied dramatically improves discrimination performance based on false negative and false positive levels.

A skilled person will appreciate that variations of the disclosed arrangements are possible without departing from the invention. For example, whilst the validation techniques are described primarily with reference to currency validation, they could equally be applied to other documents that are rich in image data, such as event tickets or passports. In addition, although the classifiers described herein are based on a parametric D² approach or a semi-parametric test, it will be appreciated that any suitable one-class classifier could be used. One-class classification techniques that could be used include the Support Vector Data Domain Description (SVDD), also known as 'support estimation', described in "Support vector domain description" by Tax, Pattern Recognition Letters 20(11-13) 1999 1191-1199, and Extreme Value Theory (EVT), which is described in the article "Novelty detection using extreme value statistics", by Roberts, IEE Proceedings on Vision, Image & Signal Processing, 146(3) (1999) 124-129. In SVDD the support of the data distribution is estimated. This avoids the difficult problem of density estimation in the case where data is sparse in high dimensional situations, whilst EVT estimates the distribution of extreme values. Accordingly, the above description of a specific embodiment is made by way of example only and not for the purposes of limitation. It will be clear to the skilled person that minor modifications may be made without significant changes to the operation described.

Claims

A method for creating a template, or data representative of such a template, for a document that has to be validated, such as a banknote, the method comprising: segmenting images of a plurality of reference documents in a like manner into a plurality of segments, and processing corresponding segments of the plurality of reference document images to determine one or more reference image segments or data representative of one or more such reference image segments.
A method as claimed in claim 1, wherein the step of processing involves classifying corresponding segments of each image to determine one or more reference segment classification parameters for each segment.
A method as claimed in claim 2 comprising defining a threshold for the reference classification parameter.
A method as claimed in claim 2 or claim 3, wherein the step of classifying comprises using a one-class classifier.
A method as claimed in claim 4, wherein the one-class classifier uses a parametric D² test or a semi-parametric test based on a mixture of Guassians, preferably employing a bootstrap.
A method as claimed in any of the preceding claims comprising optimising the step of segmenting.
A method as claimed in claim 6 comprising optimising the step of segmenting by determining a preferred number of segments.
A method as claimed in claim 6 or 7 comprising optimising the step of segmenting by determining a preferred combination of segments for use in a validation process.
A method as claimed in claim 8, wherein the preferred combination of segments is the combination that provides a minimum number of false acceptances.
A method as claimed in any of claims 6 to 8 comprising optimising the step of segmenting using a stochastic algorithm such as a genetic algorithm.
A method as claimed in claim 10 comprising using a genetic algorithm that is adapted to: (1) initialize a random population of n chromosomes; (2) perform crossover and mutation operations to create another n offspring; (3) carry out a validation operation, and calculate a fitness for each chromosome; (4) select n chromosomes with the best fitness among all the parents and offspring as the next generation, and (5) if convergent or over a preset maximum number of iteration steps then stop, otherwise go to step (2).
A method as claimed in any of the preceding claims further comprising capturing the images of the plurality of documents, preferably using an optical or thermal or magnetic imager or a combination of these.
A method as claimed in any of the preceding claims, wherein the reference documents are genuine documents, such as genuine banknotes or counterfeit documents, such as counterfeit notes.
A computer program, preferably on a data carrier or computer readable medium, for creating a reference image or data representative of a reference image for use in a validation process, the computer program having code or instructions for segmenting each one of a plurality images of reference documents such as banknotes in a like manner into a plurality of segments, and processing corresponding image segments from the plurality of images to determine a reference image segment or data representative of such a reference image segment.
A computer program as claimed in claim 14 comprising code or instructions for processing are operable to compare like segments of the plurality of images to determine a reference segment image or reference segment parameter for each segment.
A computer program as claimed in claim 14 or claim 15 wherein the code or instructions for processing comprise code or instructions for classifying each segment to determine a reference classification parameter for each segment, and preferably a threshold for the reference classification parameter.
A computer program as claimed in claim 16, wherein the code or instruction for classifying implement a one-class classifier, for example a one-class classifier that uses a parametric D² test or a semi-parametric test based on a mixture of Guassians, preferably employing a bootstrap.
A computer program as claimed in any of claims 14 to 17 comprising optimising the step of segmenting, preferably by determining a preferred number of segments and/or a preferred combination of segments for use in a validation process.
A computer program as claimed in claim 18 that has code or instructions for implementing a stochastic algorithm for optimising the step of segmenting.
A system for developing a template for a document, the system comprising means for segmenting each one of a plurality images of reference documents such as banknotes in a like manner into a plurality of segments; and means for processing like segments of the images to determine a reference image segment or data representative of a reference image segment.
A system as claimed in claim 20, wherein the means for processing are operable to classify the segment image data, and, preferably, define a threshold for the reference classification parameter.
A self-service terminal, such as an automated teller machine comprising means for segmenting each one of a plurality images of reference documents such as banknotes in a like manner into a plurality of segments; and means for processing like segments of the images to determine a reference image segment or data representative of a reference image segment.
A self-service terminal as claimed in claim 22, wherein the means for processing are operable to classify the segment image data, and, preferably, define a threshold for the reference classification parameter.
A self-service terminal as claimed in claim 22 or claim 23 comprising an imager or scanner for capturing images of the reference documents, preferably wherein the imager or scanner is an optical or thermal or magnetic imager or scanner.
A method for verifying or validating documents comprising capturing an image of the document; segmenting that captured image into a plurality of different segments; comparing one or more of the segments of the captured image, or data representative of those segments, with corresponding segments of a reference image, or data representative of those reference segments, and determining whether the document is valid or invalid based on the results of the step of comparing.
A method as claimed in claim 25, wherein the step of comparing the segments of the image with the corresponding segments of the reference image involves classifying each segment to determine a classification parameter, and comparing that classification parameter with a reference classification parameter for the corresponding segment of the reference image.
A method as claimed in claim 26 wherein the step of classifying uses a one-class classifier.
A method as claimed in claim 27, wherein the one-class classifier uses a parametric D² test, or a semi-parametric test based on a mixture of Guassians, preferably employing a bootstrap.
A method as claimed in any of claims 25 to 28 further involving comparing a plurality of the captured image segments with a corresponding plurality of the reference segments, thereby to determine the validity of each of these segments.
A method as claimed in claim 29 wherein to determine the overall validity of the document the method involves combining the validity of the plurality of the segments.
A method as claimed in claim 30, wherein the step of combining involves using a unanimous vote, in which all of the segments considered have to be valid in order to give an overall result that the document is valid.
A method as claimed in claim 30, wherein the step of combining involves using a majority vote, in which only a majority of the segments considered have to be valid in order to give an overall result that the document is valid.
A method as claimed in any of claims 25 to 32 wherein when a document is determined to be valid, the captured image or any data associated with it may be used to up-date the reference image.
A computer program, preferably on a data carrier or computer readable medium, comprising code or instructions for using a captured image of the document; segmenting that captured image into a plurality of different segments; comparing one or more of the segments of the captured image, or data representative of the one or more captured segments, with corresponding segments of a reference image, or data representative of the reference segments, and determining whether the document is valid or invalid based on the results of the step of comparing.
A computer program as claimed in claim 34, wherein the code or instructions for comparing the segments of the image with the corresponding segments of the reference image classify each segment to determine a classification parameter, and compare that classification parameter with a reference classification parameter for the corresponding segment of the reference image, preferably using a one-class classifier.
A computer program as claimed in claim 34 or claim 35 comprising code or instructions for comparing a plurality of the captured image segments with a corresponding plurality of the reference segments, thereby to determine the validity of each of these segments.
A computer program as claimed in claim 36 wherein determining the overall validity of the document involves combining the validity of the plurality of the segments, preferably using a unanimous vote, in which all of the segments considered have to be valid in order to give an overall result that the document is valid or a majority vote, in which only a majority of the segments considered have to be valid in order to give an overall result that the document is valid.
A computer program as claimed in any of claims 34 to 37 wherein when a document is determined to be valid, the captured image or any data associated with it is used to up-date the reference image or reference classification parameter.
A system, preferably a self-service terminal, such as an ATM, for verifying or validating documents, the system comprising means for capturing an image of the document; means for segmenting that captured image into a plurality of different segments; means for comparing one or more of the segments of the captured image, or data representative of the one or more captured image segments, with corresponding segments of a reference image, or data representative of the reference segments, for determining whether the document is valid or invalid based on the results of the step of comparing.
A system, preferably a self-service terminal, such as an ATM, for verifying or validating documents, the system comprising a scanner or imager for capturing an image of the document, and a processor configured to segment that captured image into a plurality of different segments; compare one or more of the segments of the captured image with corresponding segments of a reference image, and determine whether the document is valid or invalid based on the results of the step of comparing.
A method for creating a template for a document, such as a banknote, that has to be validated, or data representative of such a template, the method comprising using a scanner or imager to capture images of a plurality of reference documents or a pre-determined part of these documents, such as banknotes, and defining a reference template, or data representative of such a template using the captured images.
A method as claimed in claim 41 wherein the step of defining involves classifying the image using a one-class classifier to produce classification data for each image, and using the classification data as reference data for use in a subsequent validation process.
A method for verifying or validating documents comprising capturing an image of the document, classifying the image using a one-class classifier, comparing the classification data with corresponding reference classification data and determining whether the document is valid or invalid based on the results of the step of comparing.